Time Series Classification (TSC) problems are a rising subject of re-
search in the world of machine and deep learning. Many methods have
been developed in the last two decades over problems such as voice and
image recognition, with many everyday applications. While financial asset
price charts have been a focal example of time series, recognizing charac-
teristic patterns that may give away the future trend of an asset’s price
is a more novel method.In this paper we attempted to transform real-
valued data with SAX(Symbolic Aggregate approXimation), and created
a novel rule-based approach to extract patterns as strings of characters, as
well as an algorithm called CPC-SAX to predict unlabeled data, through
weighted distance of the characters of each string. The results show high
accuracy, for 12 characteristic chart patterns and four different time win-
dows of 15,30,45, and 60 days.The correlation between the appearance of
patterns and time windows is also highlighted. We aspire to add more
chart patterns in the labelling process and refine both the rule-based ap-
proach and distanced-based prediction of the algorithm, in future work.
(EL)
Time Series Classification (TSC) problems are a rising subject of re-
search in the world of machine and deep learning. Many methods have
been developed in the last two decades over problems such as voice and
image recognition, with many everyday applications. While financial asset
price charts have been a focal example of time series, recognizing charac-
teristic patterns that may give away the future trend of an asset’s price
is a more novel method.In this paper we attempted to transform real-
valued data with SAX(Symbolic Aggregate approXimation), and created
a novel rule-based approach to extract patterns as strings of characters, as
well as an algorithm called CPC-SAX to predict unlabeled data, through
weighted distance of the characters of each string. The results show high
accuracy, for 12 characteristic chart patterns and four different time win-
dows of 15,30,45, and 60 days.The correlation between the appearance of
patterns and time windows is also highlighted. We aspire to add more
chart patterns in the labelling process and refine both the rule-based ap-
proach and distanced-based prediction of the algorithm, in future work.
(EN)
